Monte Carlo Localization (MCL) is a common method for self-localization of a mobile robot under the assumption that a map of the environment is available. In addition to laser scanners and sonar sensors, localization approaches using vision sensors have also been recently developed with good results. In this paper we present two variations to improve the standard implementation of the MCL algorithm. The first change consists in a new strategy for the generation of particles, both at the initialization and at the resampling stage, which tries to generate new particles near the position of images in the learning dataset or in the neighborhood of particles with higher weights in the previous estimate, respectively. The second variation is related to a new approach to the estimate of the robot position, now based on two steps: clustering of particles and taking as estimate of robot position the center of the cluster, computed as a weighted sum of particle weights, with higher weight. The improved MCL algorithm described in this paper is compared with the standard MCL algorithm in terms of localization accuracy. In particular, tests were performed using local feature matching of omnidirectional images implemented on a real robot system operating in large outdoor environments with high dynamic content. Obtained results show that the localization accuracy of the improved MCL algorithm is more than twice that of the standard MCL algorithm.

This content is only available via PDF.
You do not currently have access to this content.